default-output-block.skip-main
Politics

The first government agency to trial generative AI

ACC is the first government agency to trial Microsoft's Copilot, a generative AI programme. (File image) Photo / AFP / Joe Raedle

This article was first published by RNZ.

ACC has become a chief cheerleader for the government’s push to use more artificial intelligence across public services, saying it has been a big boost in tests so far.

The compensation corporation is the first government agency to trial generative AI - Microsoft’s Copilot - in tests over the past year.

Newly released internal reports showed four fifths of its 300 users found a boost to the quality and speed of their work, as well as boosts for people with dyslexia or poor hearing.

“No other government agencies have trialled this tool ahead of ACC,” internal reports released to RNZ under the Official Information Act said.

ACC had imagined AI might boost speed and efficiency. So it set out in 2023 to “understand if those benefits can be realised ... without compromising the IP [intellectual property] and privacy of ACC clients”.

The AI operated on a fixed dataset. It did not collect information, nor did it tap into the main client record systems, so privacy risks were low.

It did not learn from the queries staff made or the information they used with it, and did not add that information to its learning banks, the reports said.

“This work is ... of great interest across the other government departments,” an ACC ethics panel’s report said.

“As one of the few government organisations who has the approval and capability to participate (given our mature technology environment), the outcome of our work on this will help inform other government organisations who may wish to also use M365 Copilot.”

The panel raised the issue of “any conflict” in Microsoft providing an AI tool to a Crown agency.

It was told the government’s chief digital officer endorsed the trial, and got approval from the government’s chief privacy officer and the Office of the Privacy Commissioner.

The two tests - first with 25 staff, then with 300 - found that along with boosts to service came gains in employee wellbeing, such as helping people with ADHD or poor hearing focus more in meetings, or those with dyslexia to revise content.

“Surveys done ... have indicated that it has been extremely beneficial,” the reports said. Getting access to it was “highly sought after”.

The corporation told RNZ it was spreading the word around.

“These findings are being openly shared with other public sector agencies and technology forums.”

Recently, the government has talked up AI for education and health purposes and begun introducing a ‘dumb’ chatbot - it cannot think for itself - to help businesses access information.

ACC clients have often expressed concern about barriers to claims and service, about what information the corporation holds on them, and what the corporation does with it.

“The data and prompts are not retained by the large learning module, nor does it use the data and prompts we provide to learn,” the ethics report said.

Some information was off-limits, such as business apps, it said.

A privacy impact assessment was also newly released.

It said the main risk of the tool was privacy breaches of personal information, whether by hacking, because it was inaccurate, or the misuse of it.

“The risks of breaching client personal information are mostly alleviated because [the AI] does not integrate with ACC’s authoritative client record systems,” the assessment report said.

“Therefore, Copilot is unable to collect, use or disclose personal information from those systems.”

It left a trail behind for easy auditing, although while the ethics panel floated the possibility of watermarking AIed documents, it was noted documents that had been spellchecked were not watermarked.

All users were being trained on how to ask useful questions and review any output.

If AI’s use was spread more broadly, the risks would rise, the documents showed.

“The primary risk is potential inherent AI bias.

“The privacy team suggests considering implementing spot checking to ensure client personal information is not directly entered into the tool.”

The report laid out a range of regular reviews to keep tabs.

AI had its limits: Te reo Māori was not a supported language on Copilot, and the assessment said any attempts at translation needed thorough review.

“Transliteration is a closer description of what the tool attempts to do.

“Computer generated translations are not encouraged, and ... qualified and recognised translators should be used.”

ACC’s Māori advisers were part of the trials.

The Official Information Act response showed ACC spent $70,000 on contractors for the trials, but it refused on commercial grounds to disclose what it had paid Microsoft for 300 year-long licences, or what it expected to pay in future.

Microsoft revenues grew to $1.2 billion in 2023, up from $1 the year before and $775m two years ago. Its profit was steady at about $40m.

Microsoft is promoting what it calls “an everyday AI companion. We are building Copilot into all our most used products and experiences”.

- RNZ